Semantic loss marks early Alzheimer's disease‐related neurodegeneration in older adults without dementia

Abstract Objective To assess progression of semantic loss in early stages of cognitive decline using semantic and letter fluency performance, and its relation with Alzheimer's disease (AD)‐specific neurodegeneration using longitudinal multimodal neuroimaging measures. Methods Change in verbal fluency was analyzed among 2261 non‐demented individuals with a follow‐up diagnosis of no mild cognitive impairment (MCI), amnestic MCI (aMCI), non‐amnestic MCI (naMCI), or incident dementia, using linear mixed models across 4 years of follow‐up, and relations with magnetic resonance imaging (MRI; n = 1536) and 18F‐fluorodeoxyglucose brain positron emission tomography (18F‐FDG‐PET) imaging (n = 756) using linear regression models across 2 years of follow‐up. Results Semantic fluency declined—fastest in those at higher risk for AD (apolipoprotein E [APOE] e4 carriers, Clinical Dementia Rating score of .5, aMCI, or incident dementia)—while letter fluency did not except for those with incident dementia. Lower baseline semantic fluency was associated with an increase in white matter hyperintensities and total mean cortical thinning over time, and regionally with less hippocampal volume as well as more cortical thinning and reduced 18F‐FDG‐PET uptake in the inferior parietal lobule, entorhinal cortex, isthmus cingulate, and precuneus–posterior cingulate area. In contrast, baseline letter fluency was not associated with change in total nor regional neurodegeneration. Whole‐brain neurodegeneration over time was associated with faster decline in both fluencies, while AD‐specific regions were associated with a faster rate of decline in semantic but not letter fluency. Interpretation This study provides strong evidence of distinctive degeneration of semantic abilities early on in relation to both cognitive decline and AD‐specific neurodegeneration.


INTRODUCTION
The preclinical phase of Alzheimer's disease (AD) is not only marked by amyloid and tau accumulation and neurodegeneration, 1 but also by subtle cognitive changes years before a clinical diagnosis can be established. 2,3 A diagnostic marker of AD in clinical practice, and supported by observations in research, is a diverging performance pattern of semantic fluency versus letter fluency 4,5 -generating as many words within time limits that start with a specific category or letter, respectively. Cognitively normal individuals typically perform better on semantic fluency than letter fluency, 6,7 while individuals with manifest clinical AD often show greater impairment in semantic fluency compared to letter fluency. [8][9][10] The reversal of this pattern over the course of the preclinical to clinical AD process is due to a loss of semantic memory, which is one of the first cognitive domains to become impaired in addition to episodic memory. 11,12 How the discrepancy between letter and semantic fluency relates to neurobiological change over time in the early stages of AD, however, is relatively unexplored. The cortical signature of neurodegeneration in early stages of AD includes medial-temporal and temporal-parietal regions. 13,14 In the left hemisphere, these regions are particularly associated with semantic processing abilities. [15][16][17] Correspondingly, semantic fluency has been linked to temporal-parietal as well as frontal regions, while letter fluency is considered to be relatively confined to inferior frontal regions important for executive functioning. 7,18 The discrepancy between semantic and letter fluency in clinical AD has been ascribed to semantic deficits that are mediated by neurodegeneration of temporal-parietal regions. 4,19 Identifying how the discrepancy in semantic versus letter fluency develops in a preclinical phase, and how it relates to change over time in AD markers of neurodegeneration, will provide better insight into the potential predictive value of this discrepancy in the earliest stages of the AD process. Potential implementations include use of this knowledge to refine the definition of high-risk individuals for clinical trials aimed at intervention.
Additionally, knowledge about the degree of discrepancy and its development over time may be used as a measure of progression in disease-modifying interventions.
This study aimed to (1)

Participants
Participants were drawn from the Memento cohort, an ongoing multicenter prospective study on AD and related disorders of 2323 individuals recruited from French memory clinics between 2011 and 2014.
Participants were screened at inclusion to be non-demented (Clinical Dementia Rating [CDR] of ≤.5). All examinations in Memento followed standardized procedures, including neuropsychological assessment, brain MRI, and 18 F-FDG-PET imaging. The cohort, recruitment, design, and procedures are described in detail elsewhere. 20 In the present study, individuals were excluded from analysis if they were missing semantic fluency (n = 15), letter fluency (n = 6), or both (n = 33) at baseline; if they were diagnosed with prevalent dementia at baseline (n = 1); or if they were missing demographic information Semantic and letter fluency should be included as basic tasks in every cognitive assessment in aging research.
(completed high school), or high (completed a higher-level diploma).
Participants were genotyped for apolipoprotein E (APOE; n = 2146) and categorized as APOE e4 carriers based on the presence of at least one e4 allele. 20 22,23 The cognitive status at last follow-up provides future diagnostic information about the trajectory of an individual when assessing cognitive change over time. 24,25 For example, when cognitive status at last follow-up is dementia, this information confirms that an individual had pre-clinical dementia at baseline.
The project was approved by the ethics committee ("Comité de Protection des Personnes Sud-Ouest et Outre Mer III") and followed the guidelines of the Declaration of Helsinki. All participants gave written consent.

Cognitive measures
An extensive neuropsychological battery was administered annually, including tasks of memory, language, praxis, visuospatial skills, and executive functioning. 20

Statistical analysis
Participant characteristics were extracted using descriptive statistics.
Differences in semantic and letter fluency performance were analyzed with general linear models.
To analyze the change of letter and semantic fluency performance over time, linear mixed models were used. Models included either semantic or letter fluency as an outcome, with time in study (in years, starting at 0), age at baseline, an interaction term of time in study and age at baseline (to account for age-related differential decline 32,33 ), sex, education, and practice effect as fixed factors, as well as a random intercept and random slope: Models with an interaction term of time in study and age at baseline fitted better than those without this interaction for both semantic (Akaike information criterion [AIC] with 56609.01 vs without 56615.98) and letter fluency (AIC with 52805.31 vs without 52811.59). Semantic and letter fluency were standardized by subtracting the test's mean score of the study sample at baseline from each individual's score, and dividing by the study sample's standard deviation at baseline. To report effects of fluency performance at baseline (ie, intercept) and across time (eg, slope), age was centered and sex and education were treated as covariates as opposed to factors to reflect performance of the average participant in all models. Practice effects, often representing reduced anxiety on successive testing occasions, were modeled using an indicator variable being the square root of the number of prior testing occasions. 34 This variable represents that practice effects are present throughout follow-up, but that the largest effects occur after the first exposure and gradually diminish across follow-up visits.
The intraclass correlation coefficient (ICC; variance of a random effect/total random variance) was calculated for intercept and slope in the overall models of semantic and letter fluency to indicate how much variance was explained by the random effects. We have also reported the correlation coefficient between intercept and slope (I-S corr) for the overall models of semantic and letter fluency, which reflects the degree to which individuals with a higher baseline decline at a faster rate than those with a lower baseline.
In separate models, analyses were stratified by cognitive status at follow-up (no MCI, aMCI, naMCI, incident dementia), education (low, middle, high), sex (men/women), CDR score at inclusion (0/.5), and APOE status (e4+/e4-). Differences in slopes across groups were formally tested with an interaction of each variable with time in study; we did not include a triple interaction of years in study, group, and age in these models because the interaction was non-significant and worsened model fit for all models. Interactions were performed with the following reference groups: no MCI and naMCI for cognitive status, low and middle for education, men for sex, 0 for CDR score, and e4-for APOE status.
To analyze the relationship of baseline performance and rate of change (ie, slope) of semantic versus letter fluency with longitudinal multimodal neuroimaging measures, multiple linear regression models were used. Individual rates of change were extracted from the previously estimated linear mixed models (ie, individual estimates for time in study regressed on fluency performance), adjusted for age, age x time in study, sex, education, and practice effect. The multiple linear regression models analyzed whether semantic or letter fluency (baseline and rate of change) were predictors of follow-up neuroimaging, which was autoregressed on neuroimaging at baseline. Models of baseline performance adjusted for age, sex, education, and the interaction of magnet strength (1.5T or 3T) with the manufacturer (Siemens, Phillips, or GE Healthcare) to account for center-specific differences. Models of rate of change adjusted for the interaction of magnet strength with the manufacturer, as the slopes were previously adjusted for age, sex, education, and practice effect. Differences across groups based on cognitive status at follow-up were formally tested with an interaction of cognitive status with fluency performance. Subsequent models added the other fluency measure as a covariate to analyze the discrepancy between letter and semantic fluency. 2 Autoregression accounts for baseline levels but may sometimes induce bias; 35 therefore, the same analyses were run in models using difference scores between baseline and follow-up neuroimaging measurements (however, this method does not account for the initial level of the dependent variable). Both options to model change yielded similar patterns, and below we report the results of autoregression.
The multiple linear regression models were performed for wholebrain measures of neurodegeneration (ie, WMH, total mean cortical thickness, and total 18 F-FDG-PET SUVR) and for regional measures to investigate localization of effects (ie, hippocampal volume, cortical thickness of the entorhinal cortex, parahippocampal gyrus, and isthmus cingulate, and SUVR of the precuneus-posterior cingulate area and inferior temporal area).

Participant characteristics
Demographic information, cognitive status, CDR score, and APOE e4 status for the overall sample (N = 2261), as well as for the subsets who underwent MRI (n = 1542) and 18     Note. 1 Significance of slope effects: * P < .05, ** P < .01, *** P < .001. Models are adjusted for age at baseline, age at baseline x time in study, sex, and educationunless sex or education was used to stratify-and models of letter fluency were additionally adjusted for time in study x time in study Abbreviations: aMCI, amnestic MCI; APOE, apolipoprotein E; CDR, Clinical Dementia Rating; ICC, intraclass correlation coefficient; I-S Corr, intercept-slope correlation; MCI, mild cognitive impairment; naMCI, non-amnestic MCI; SE, standard error.

Prediction of follow-up neuroimaging by baseline fluency performance
Lower baseline semantic fluency, but not letter fluency, was associated with more neurodegeneration over time in whole-brain neuroimaging measures of total mean cortical thickness and WMH volume, but not total SUVR on 18 F-FDG-PET (Table 3 and Figure 3). These patterns for whole-brain measures of total mean cortical thickness and WMH, but did not differ across cognitive status groups in ROI analyses (Table   S1 in supporting information). Table 3 presents the relationships of follow-up whole-brain neurodegeneration and ROIs with change in fluency performance ( Figure S1 in supporting information). Faster rate of decline in both semantic and letter fluency was related to more total mean cortical thinning over time in unadjusted and adjusted models. This relationship for semantic fluency was stronger in aMCI and incident dementia groups compared to the naMCI group, while the relationship with letter fluency was stronger in the naMCI group compared to the no MCI group (Table   S1). Faster rate of decline in both semantic and letter fluency was also related to an increase in WMH volume over time in unadjusted models and attenuated for semantic fluency in adjusted models; the strength of the relationship did not differ across groups of cognitive status at last follow-up. Change in whole-brain 18

DISCUSSION
We demonstrated the progression of semantic impairment at baseline and semantic decline over time in early stages of cognitive decline and whole-brain and regional neurodegeneration measures showed that the relationship between the rate of decline in semantic fluency (but not letter fluency) and more AD-related neurodegeneration over time was stronger in the aMCI than naMCI group-aMCI is known to have a high likelihood of progressing to AD whereas naMCI has a higher likelihood of progressing to non-AD dementia. 36  The robustness of this pattern is shown in every stratum we analyzed (Table 2). That the pattern is stronger in those who are at higher risk for AD dementia based on higher CDR score, APOE e4 positivity, or a cognitive status of aMCI or incident dementia at last follow-up (as opposed to subjective complaints only) corresponds with a prevalence study on the discrepancy of the verbal fluency pattern in AD by Sherman and Massman. 38  Classically, analyses of cluster and switches have already been shown to add above and beyond total number of items in detecting MCI 40,41 and very mild AD, 42 and usage of psycholinguistic variables has demonstrated its value in distinguishing cognitively normal individuals at increased genetic risk of AD dementia by virtue of carrying an APOE e4 allele. 43 More research in this area should be encouraged, as the present study adds more evidence to the growing body of research on semantic breakdown in the very early stages of AD. This knowledge may be used as a sensitive clinical marker in preclinical AD. 44 The emerging discrepancy in semantic versus letter fluency in the early stages of AD was related to the neurodegenerative processes of the disease, including hippocampal atrophy, whole-brain and regional cortical thinning, regional 18 F-FDG-PET SUVR, and progression of small vessel disease as measured by WMH. 45 The discrepancy in semantic versus letter fluency may have future potential as an inexpensive and easy-to-administer proxy for AD-related neurodegenerative processes otherwise detected on expensive and time-consuming brain scans, but this application for clinical use should be further investigated.
Semantic fluency is considered to call on two cognitive abilities, namely semantic processing and executive functioning, and letter fluency is considered to call predominantly on executive functioning. 46 By controlling for letter fluency when analyzing semantic fluency, and vice versa, the effects of the semantic processing versus executive functioning components can be isolated to a certain extent, 2 to show that these two fluency tasks tap distinctive cognitive processes. 41 Our results display the hybrid character of both tasks in their cortical signature: The ROI analyses confirm previous findings of letter fluency mediating inferior frontal regions and semantic fluency mediating both inferior frontal and temporal-parietal regions. 7,18 The impairment in semantic fluency in clinical AD is thought to be mediated by neurodegeneration of temporal-parietal regions. 4,19 Our study showed that indeed, Limitations of this study include neuroimaging at no more than two time points in our data. Nonetheless, the Memento cohort is a cohort study in progress, and one more neuroimaging session is planned for every participant during future follow-up visits. The Memento cohort is a large and unique collection of individuals who visited a memory clinic, as there was no exclusion based on the cognitive domain of complaints or impairment (eg, isolated non-memory deficits). This inclusive and non-selective approach resulted in a real-life cohort that represents the whole scope of individuals that come to a memory clinic, which is important for generalization purposes. However, due to its nature of being a memory clinic cohort, another limitation is that the Memento cohort only includes individuals with subjective or objective cognitive impairment. This condition prohibited us from comparing the fluency trajectories of these individuals and the relationships of their fluency performance with neurodegeneration to those in cognitively healthy individuals without any subjective or objective cognitive impairment, who may show different trajectories or cognitive-neural associations.
Nonetheless, a strength of the Memento cohort is that it includes sizeable groups of individuals with amnestic and naMCI (in addition to no MCI and incident dementia), which enabled us to compare analyses across these two groups that are known to are more likely to progress to AD versus non-AD dementia, respectively. 36 Multiple cross-sectional studies on semantic and letter fluency that compared different neurodegenerative diseases have shown dissociative patterns in the two verbal fluency tasks, in relation to regional cortical or subcortical damage. 50,51 In our study, the naMCI group, which is thought to be at higher risk for non-AD dementia, showed different and letter fluency can be used as a specific marker (in addition to a sensitive marker) for AD compared to other dementias. 52 In sum, the non-invasive, low-cost, classic, and easy-to-administer cognitive measure of semantic fluency, and its diverging performance from letter fluency, may aid preclinical detection of AD. This study attests to the value of including semantic and letter fluency as basic tasks in every cognitive assessment in aging research. The discrepancy in semantic and letter fluency performance may be useful in diagnostic prediction of future AD-related neurodegeneration across multimodal neuroimaging.

ACKNOWLEDGMENTS
The MEMENTO cohort is funded by the Fondation Plan Alzheimer

CONFLICTS OF INTEREST
The authors have no relevant conflicts of interest or financial or other nonprofessional benefits to disclose that could bias the authors in the conduct of the reported work.